Three-dimensional representation method for fracture distribution based on intelligent fusion of multi-source data

Abstract

This invention discloses a three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data, relating to the field of unconventional oil and gas development. The method first establishes a fracturing model based on field data, simulating the horizontal well fracturing process and microseismic events, and constructs a multi-source database including reservoir properties, pumping design, wellbore-natural fracture images, and microseismic data. Subsequently, the data is preprocessed to generate two-dimensional microseismic feature maps and wellbore-natural fracture grayscale maps. Based on this, a fracturing fracture distribution prediction model based on multi-source data feature fusion is constructed, introducing an attention mechanism to achieve dynamic weighting of cross-modal features, and the model is trained. Finally, using the validated model and field-acquired data, the fracture distribution is predicted and a three-dimensional reconstruction map is generated. This invention improves prediction accuracy and robustness through deep fusion of multi-source data, achieving high-precision, interpretable, and visual prediction of fracturing fracture distribution.

Description

Technical Field

[0001] This invention relates to the field of unconventional oil and gas field development technology, specifically to a three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data. Background Technology

[0002] In the development of unconventional oil and gas fields, accurately characterizing the fracture distribution within the reservoir is crucial for fracturing scheme design and improving oil recovery. However, current methods for characterizing fracturing fracture distribution face numerous technical challenges, as follows:

[0003] First, processing multi-source data is complex;

[0004] Since fracture characterization involves information from multiple sources such as geological reservoir parameters, fracturing operation parameters, and seismic monitoring data, and these data are characterized by diverse data formats, different scales and resolutions, and large spatiotemporal differences, they are difficult to integrate effectively and often require simplification. This leads to insufficient utilization of data information and increases the uncertainty of the three-dimensional characterization of fracturing fracture distribution.

[0005] Second, the accuracy of characterization needs to be improved;

[0006] Currently used fracture characterization methods are mostly based on a single data source or empirical model. For example, relying solely on geological modeling and numerical simulation to obtain fracture distribution maps or solely on inverting fracture networks from microseismic events is highly susceptible to data errors. Furthermore, the lack of comprehensive consideration of multiple factors results in low accuracy in predicting the distribution of hydraulic fracturing fractures, frequently leading to missed or incorrect predictions, and failing to meet the needs of detailed fracture characterization in the field.

[0007] Third, feature extraction is limited;

[0008] Traditional three-dimensional characterization methods for fracturing fracture distribution often focus on specific data features, failing to comprehensively depict the multiple factors influencing fracture formation and propagation. For example, relying solely on static geological parameters is insufficient to reflect the dynamic fracturing behavior during fracturing; and relying solely on microseismic dynamic information ignores static background factors such as lithology and stress fields. In other words, the lack of collaborative extraction of multimodal features in existing technologies results in insufficient generalization ability and robustness of fracturing fracture distribution prediction models, making them unsuitable for fracture prediction requirements under different reservoir conditions.

[0009] Fourth, the results have limited dimensions of representation;

[0010] Currently, the characterization results of fracturing fracture distribution are mostly presented in the form of two-dimensional planar diagrams or local cross-sections, which are insufficient in the three-dimensional depiction of fracture spatial distribution. This makes it difficult for engineers to intuitively understand the fracture network structure of the entire reservoir. The lack of a complete three-dimensional fracture distribution characterization directly affects the accuracy of reservoir exploitation decisions and construction efficiency.

[0011] Therefore, there is an urgent need to propose a three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data, which can improve the characterization accuracy of fracturing fracture distribution by integrating multi-source heterogeneous data and achieve efficient and accurate characterization of the spatial distribution of fracturing fractures. Summary of the Invention

[0012] This invention aims to solve the above-mentioned problems and provides a three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data. It establishes a fracturing fracture distribution prediction model that can fuse features from multiple sources of data. By extracting reservoir attribute features, pumping embedding features, microseismic embedding features, and wellbore-natural fracture distribution image features, channel attention is introduced to perform cross-modal dynamic weighted fusion. By encoding and decoding the fused features of multi-source data, a fracture distribution probability map is obtained. The fracturing fracture distribution prediction model is used to predict the actual fracture distribution in the reservoir, achieving efficient spatial distribution and accurate characterization of fracturing fractures.

[0013] To achieve the above objectives, the present invention adopts the following technical solution:

[0014] A three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data includes the following steps:

[0015] Step 1: Establish a fracturing model based on the fracturing construction site data, and use the fracturing model to simulate the fracturing construction process and microseismic events of horizontal wells in the reservoir. Obtain reservoir attribute datasets, pumping design datasets, wellbore-natural fracture distribution image sets, fracturing fracture distribution image sets, and microseismic event datasets to construct a multi-source database for fracturing fracture characterization.

[0016] Step 2: Preprocess the multi-source data in the multi-source database for fracturing fracture characterization to obtain two-dimensional microseismic feature maps of each reservoir layer and grayscale images of wellbore-natural fracture distribution.

[0017] Step 3: Construct a fracturing fracture distribution prediction model based on multi-source data feature fusion. The fracturing fracture distribution prediction model includes a reservoir attribute data processing module, a pumping design data processing module, a microseismic event data processing module, an image convolution module, a multi-source data dynamic weighted fusion module, and a fracture distribution prediction module.

[0018] Step 4: Use multi-source data from the multi-source database of fracturing fracture characterization to train and validate the fracturing fracture distribution prediction model based on multi-source data feature fusion, and obtain the validated fracturing fracture distribution prediction model.

[0019] Step 5: Based on the reservoir attribute data and pumping design data collected on site, input the wellbore-natural fracture distribution grayscale image constructed based on the on-site geological data and perforation scheme and the simulated microseismic event data into the validated fracturing fracture distribution prediction model. Use the fracturing fracture distribution prediction model to predict the fracturing fracture distribution and generate a three-dimensional fracture distribution reconstruction map.

[0020] Preferably, step 1 includes the following sub-steps:

[0021] Step 1.1: Obtain multiple fracturing operation site data, including reservoir data, fracturing design schemes and well completion data, and construct a reservoir attribute dataset, a pumping design dataset and a wellbore-natural fracture distribution image set;

[0022] The reservoir data includes multiple reservoir physical property parameters, namely porosity, permeability, water saturation, reservoir pressure, triaxial principal stress, Young's modulus, and Poisson's ratio; the fracturing design scheme includes multiple fracturing construction characteristic parameters, namely pumping rate between fracturing horizontal well sections, fracturing fluid type, proppant type, and proppant concentration; the completion data includes wellbore perforation images and natural fracture distribution images of each fracturing section in the fracturing horizontal well, used to determine the wellbore-natural fracture image;

[0023] Step 1.2: Construct fracturing models in the fracture propagation numerical simulator based on the data from each fracturing construction site to obtain multiple fracturing models under different geological conditions and fracturing design schemes. Perform fracturing construction simulation for each fracturing model to generate fracturing fractures, obtain fracturing fracture distribution images, and construct a fracturing fracture distribution image set.

[0024] Step 1.3: When simulating fracturing construction based on each fracturing model, obtain the spatial position of the fracture mesh element and the shear displacement on the fracture mesh element in each fracturing model, determine the seismic moment, and use the preset microseismic data generation program to generate microseismic events, obtain microseismic event data including the three-dimensional spatial position and magnitude of microseismic events, and establish a microseismic event dataset.

[0025] Step 1.4: Construct a multi-source database for characterizing fracturing fractures based on reservoir attribute datasets, pumping design datasets, wellbore-natural fracture distribution image sets, fracturing fracture distribution image sets, and microseismic event datasets.

[0026] Preferably, in step 1.3, a shear displacement threshold is set based on the minimum magnitude, shear modulus, and minimum fracture zone area in the fracturing model. for:

[0027] ;

[0028] In the formula, The seismic moment; Shear modulus; The minimum area of ​​the fractured region;

[0029] Using a breadth-first search algorithm, models with shear displacements greater than a shear displacement threshold are selected from the fracturing model. Given connected components, calculate the area of ​​each connected component. centroid location and mean shear displacement ;

[0030] For each connected block, calculate the seismic moment of the connected block based on its shear displacement and area. Moment and Magnitude The calculation formula is:

[0031] ;

[0032] ;

[0033] In the formula, This represents the average shear displacement. It is a logarithmic function with base 10;

[0034] The area, centroid location, average shear displacement, seismic moment, and moment magnitude of each connected block are saved to the crack propagation mesh file to form multiple microseismic event data and construct a microseismic event dataset.

[0035] Preferably, step 2 includes the following sub-steps:

[0036] Step 2.1: Normalize the reservoir physical property parameters in the reservoir attribute dataset and the fracturing operation characteristic parameters in the pumping design dataset.

[0037] Step 2.2: Preprocess the image data in the wellbore-natural fracture distribution image set and the hydraulic fracture distribution image set. For each wellbore hole image, each natural fracture distribution image and each hydraulic fracture distribution image, first perform binarization according to the preset grayscale threshold, then perform grayscale mapping, and then crop to the specified size to obtain the wellbore hole grayscale image, the natural fracture distribution grayscale image and the hydraulic fracture distribution grayscale image, and use the hydraulic fracture distribution grayscale image as the label;

[0038] Step 2.3: Preprocess the data of each microseismic event in the microseismic event dataset;

[0039] For each microseismic event data point, after spatial transformation and layer division based on the three-dimensional spatial location of the microseismic events, a two-dimensional spatial grid with the same size as the fracturing fracture distribution image is constructed. This two-dimensional spatial grid includes three types of feature channels: microseismic event density channel, magnitude-weighted energy channel, and... Value distribution channel, used to extract microseismic event density, magnitude-weighted energy, and... Value distribution;

[0040] For each layer, using the spatial distribution of microseismic events as input to a two-dimensional spatial grid, a continuous rasterized feature image is generated using the Gaussian kernel density estimation method. Microseismic event density, magnitude-weighted energy, and other parameters are then extracted. The values ​​are distributed and normalized to limit the range of microseismic event density and magnitude-weighted energy, resulting in two-dimensional microseismic feature maps at each layer.

[0041] Step 2.4: The grayscale images of wellbore openings and natural fracture distribution at each layer are stitched together to generate a grayscale image of wellbore-natural fracture distribution. The grayscale image of wellbore-natural fracture distribution has two feature channels, namely the natural fracture channel and the wellbore opening location channel, which are used to extract the locations of natural fractures and wellbore openings.

[0042] Preferably, in step 2, the microseismic event density channel is set as follows:

[0043] ;

[0044] in,

[0045] ;

[0046] In the formula, For the first Microseismic event density of layers in a two-dimensional spatial grid; This refers to the sequence number of the microseismic event; This represents the total number of microseismic events. It is a two-dimensional Gaussian kernel function; The x-coordinate of the two-dimensional spatial grid; The ordinate of the two-dimensional spatial grid; Two-dimensional spatial grid Place and No. Micro-earthquake events in The difference in the index of the direction; Two-dimensional spatial grid Place and No. Micro-earthquake events in The difference in the index of the direction; is the base of the natural exponential function; For the first Horizontal index of microseismic events in a two-dimensional spatial grid; For the first Vertical index of microseismic events in a two-dimensional spatial grid; It is the spatial smoothing coefficient;

[0047] The magnitude-weighted energy channel is configured as follows:

[0048] ;

[0049] In the formula, For the first Magnitude-weighted energy of the layer; This is the magnitude-energy weighting coefficient; For the first The magnitude of a micro-earthquake event;

[0050] Set a partial sliding window to determine the... The value distribution channel is:

[0051] ;

[0052] In the formula, These are local frequency-magnitude distribution parameters; The average magnitude of microseismic events within the local sliding window; The minimum magnitude within the local sliding window; The intervals are for classifying earthquake magnitudes.

[0053] Preferably, the reservoir attribute data processing module is used to process reservoir attribute data, first processing the input reservoir attribute data... Dimensions converted ,in, , , , The dimensions of the reservoir attribute data are represented by the number of channels, depth, height, and width. Then, multi-layer convolution processing is applied to the dimensionally transformed reservoir attribute data. Convolutional layers extract the basic features of reservoir attribute data. After ReLU activation layers enhance the nonlinear expressive power, the data is then processed sequentially through three identical layers. After the convolutional layer extracts deep features, it utilizes... The convolutional layer performs channel compression and outputs reservoir attribute features.

[0054] The pumping design data processing module is used to process pumping design data. First, it converts the input pumping design data to frequency domain data using a Fast Fourier Transform. Then, it sequentially inputs the frequency domain data into three identical... After extracting local spectral features from the convolutional layer, then... Convolutional layers perform feature integration and compression, outputting pumped embedded features;

[0055] The microseismic event data processing module is used to process microseismic event data, first obtaining the three-channel feature tensor from the input two-dimensional microseismic feature map. For each channel feature tensor, the channel feature tensors are input sequentially into... After extracting local features from convolutional layers and ReLU activation layers, then... Convolutional layers perform channel compression to output microseismic embedding features. ;

[0056] The image convolution module is used to process grayscale images of wellbore-natural fracture distribution. It includes three sequentially connected image convolution units with identical structures. Each image convolution unit consists of a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The first image convolution unit uses a convolutional layer with 2 input channels and 32 output channels. Convolutional layers are used to extract local spatial features; the second image convolutional unit uses 32 input channels and 32 output channels. Convolutional layers are used to extract deep semantic features; the third image convolutional unit uses 32 input channels and 2 output channels. Convolutional layers are used for channel compression and outputting image features;

[0057] The multi-source data dynamic weighted fusion module is used to obtain multi-source fusion feature tensors, including a splicing layer, a parallel max pooling layer and an average pooling layer, and sequentially set... Convolutional layer, ReLU activation layer, Convolutional layers and sigmoid activation layers automatically learn the importance weights of various features in crack prediction by introducing an attention mechanism;

[0058] The fracture distribution prediction module is used to predict the fracture distribution in the reservoir. It consists of multiple parallel and identical improved U-Net structures. The number of improved U-Net structures is set according to the number of layers. Each improved U-Net structure includes an encoder, a decoder, a channel attention module, and an output layer. The encoder includes multiple encoding layers, each of which comprises sequentially connected... Convolutional layers, batch normalization layers, ReLU activation layers, and layers using... A window-downsampled max-pooling layer; the decoder includes multiple decoding layers, each decoding layer containing two identical and sequentially connected decoding convolutional units, the decoding convolutional unit including... Convolutional layers, ReLU activation layers and The convolutional layer, through upsampling and skip connections, recovers spatial resolution layer by layer, used to restore spatial resolution and detail information; the channel attention module runs through the encoder's encoding process and the decoder's decoding process, used to enhance the model's ability to identify crack development regions; the output layer is equipped with... Convolutional layers and sigmoid activation layers are used to output a crack distribution probability map.

[0059] Preferably, when the fracturing fracture distribution prediction model is applied to fracture distribution prediction, it includes the following sub-steps:

[0060] Step 3.1: Extract reservoir attribute features from each reservoir attribute data in the reservoir attribute dataset using the reservoir attribute data processing module, and then extract pumping embedding features from each pumping design data in the pumping design dataset using the pumping design data processing module. Add the extracted reservoir attribute features and pumping embedding features element by element using the element-wise addition method, and use bilinear interpolation upsampling to obtain a fused feature map. The fused feature map is provided with a reservoir-pumping fused feature channel.

[0061] Step 3.2: Use the image convolution module to extract the image features of each wellbore-natural fracture distribution grayscale image in the wellbore-natural fracture distribution image set, and use the microseismic event data processing module to extract the microseismic embedding features of each microseismic event data in the microseismic event dataset.

[0062] Step 3.3: Use the multi-source data dynamic weighted fusion module to perform feature fusion on the extracted fusion feature map, microseismic embedding features and image features to obtain a multi-source fusion feature tensor labeled with the hydraulic fracture distribution image;

[0063] The multi-source data dynamic weighted fusion module first concatenates the input fused feature map, microseismic embedding features, and image features to obtain the input feature quantity. Then, channel attention is applied to the input feature quantity, and the input feature quantity is input into the average pooling layer and the max pooling layer, respectively. After average pooling and max pooling processing, the input is then input into the sequentially connected layers. Convolutional layer, ReLU activation layer, Feature mapping is performed in the convolutional layer to obtain average pooling mapping features and max pooling mapping features. The two features are fused and then input into the Sigmoid function layer for normalization to obtain channel weight coefficients. The channel weight coefficients are multiplied and weighted by the input feature values ​​channel by channel to obtain the multi-source fusion feature tensor. The fracturing crack distribution image is used as the label of the multi-source fusion feature tensor.

[0064] Step 3.4: Input the multi-source fusion feature tensor into the fracture distribution prediction module, and use the fracture distribution prediction module to extract the spatial features from the multi-source fusion feature tensor to predict the fracture distribution in the reservoir.

[0065] Preferably, step 3.4 includes the following sub-steps:

[0066] Step 3.4.1: After normalizing the multi-source fusion feature tensor, input it into the encoder of the improved U-Net structure. The encoder then processes the data and outputs a set of feature matrices. , ,in, , , , These are all feature matrices extracted from each coding layer of the encoder;

[0067] Step 3.4.2: Calculate the encoder output feature matrix set respectively. The channel attention weights of each feature matrix are calculated using the following formula:

[0068] ;

[0069] In the formula, For the set of characteristic matrices The index of the feature matrix; This is the weighted feature matrix; For the first The feature matrix output by each coding layer; For the first Channel weight matrix of each feature matrix; Weighted operation for the channel;

[0070] Step 3.4.3: Use the decoder to process the feature matrix set respectively. Decode each feature matrix in the encoder, using the feature matrix output from the deepest layer. As input features for the decoder, each decoding layer first upsamples the input features using bilinear interpolation, and then compares them with the feature matrix output by the corresponding coding layer. The channels are skipped and spliced, then refined and channel-adjusted by the decoding convolutional units within the decoding layer. Finally, the decoded features are obtained by decoding layer by layer in the decoder, proceeding from deep to shallow. ;

[0071] Step 3.4.4: Decode the features obtained by the decoder. Input to the output layer, via After the convolutional layer compresses its channels, the data is input into the Sigmoid activation layer to generate a crack distribution probability map. The value range of each pixel in the crack distribution probability map is [value range missing]. , is used to represent the probability of crack development at a pixel.

[0072] Preferably, step 4 includes the following sub-steps:

[0073] Step 4.1: Randomly allocate various types of data from each dataset in the multi-source database for fracturing fracture characterization to the training set and validation set according to a preset ratio;

[0074] Step 4.2: Set the initial learning rate and accuracy values ​​for the fracturing fracture distribution prediction model;

[0075] Step 4.3: Input various types of data from the training set into the fracturing fracture distribution prediction model, use the training set to train the fracturing fracture distribution prediction model to predict the fracture distribution in the reservoir, and obtain the comprehensive loss function value of the fracturing fracture distribution prediction model.

[0076] The comprehensive loss function Will loss function and The loss function is coupled and set as follows:

[0077] ;

[0078] In the formula, , All are loss weighting coefficients; For focus loss function; for coefficient; For label images; The prediction results are from the fracturing fracture distribution prediction model.

[0079] Step 4.4: Compare the comprehensive loss function value of the fracturing fracture distribution prediction model with the preset accuracy value. If the comprehensive loss function value has reached the preset accuracy value, the training of the fracturing fracture distribution prediction model ends; otherwise, the learning rate of the fracturing fracture distribution prediction model is adjusted, and the process returns to step 4.3.

[0080] Step 4.5: Input various types of data from the validation set into the fracturing fracture distribution prediction model, validate the fracturing fracture distribution prediction model using the validation set, obtain the comprehensive loss function value of the fracturing fracture distribution prediction model, and compare it with the preset accuracy value. If the comprehensive loss function value has reached the preset accuracy value, the validated fracturing fracture distribution prediction model is obtained; otherwise, adjust the learning rate of the fracturing fracture distribution prediction model and return to step 4.3 to continue training the fracturing fracture distribution prediction model.

[0081] Step 4.6: Obtain the validated fracturing fracture distribution prediction model.

[0082] Preferably, in step 5, based on the reservoir attribute data and pumping design data collected on-site, the wellbore-natural fracture distribution grayscale image constructed based on on-site geological data and perforation scheme and the simulated microseismic event data are input into the validated fracturing fracture distribution prediction model, and the fracturing fracture distribution prediction model is used to predict the fracture distribution probability map.

[0083] According to the preset threshold, the crack distribution probability map is divided into non-crack areas and crack development areas. Pixels with probability values ​​less than the preset threshold are judged as non-crack areas and their pixel values ​​are set to 0. Pixels with probability values ​​not less than the preset threshold are judged as crack development areas and their pixel values ​​are set to 1, thus generating a crack prediction image.

[0084] The bilinear interpolation algorithm was used to spatially register the generated fracture prediction images with the original formation model of the actual reservoir, and the fracture prediction images of each layer were stacked along the vertical depth direction of the actual reservoir to obtain a three-dimensional fracture distribution reconstruction map.

[0085] The beneficial technical effects brought about by this invention are as follows:

[0086] (1) The present invention proposes a three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data. This method integrates multi-source data such as reservoir physical property parameters, fracturing construction characteristic parameters, wellbore-natural fracture distribution images, and microseismic events. It utilizes the fracture distribution prediction module in the fracturing fracture distribution prediction model to comprehensively learn multi-modal information and fully capture the influencing factors of fracture formation. In particular, it introduces microseismic event density, magnitude-weighted energy, and... The value distribution enhances the characterization of the dynamic response of hydraulic fracturing. Compared with existing reservoir fracture distribution prediction methods that use a single data source, the method of this invention achieves accurate characterization of reservoir fracture distribution and realistically restores the development of hydraulic fracturing fractures by fusing multi-source data, using reservoir attribute data as constraints and microseismic data as the distribution.

[0087] (2) The present invention proposes a three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data. A dynamic weighted fusion module for multi-source data is set up in the fracturing fracture distribution prediction model. This module employs an attention mechanism to assign adaptive weights to data from different sources, achieving dynamic weighted fusion of heterogeneous multi-source data. Simultaneously, the dynamic weighted fusion module can autonomously adjust the contribution of each channel feature in the fracture distribution prediction result based on the importance differences between reservoir attribute data, pumping design data, microseismic event data, and image features. This effectively avoids biases caused by manually setting weights. Furthermore, the adaptive feature fusion strategy ensures that key information is not overwhelmed, improving the robustness of the fracturing fracture distribution prediction model to complex inputs and guaranteeing the reliability of the prediction results.

[0088] (3) The present invention proposes a three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data. Based on the predicted fracture distribution probability map of each layer of the reservoir, a three-dimensional fracture distribution reconstruction map of the reservoir is reconstructed by superimposing multiple layers. This method can intuitively show the spread and connection of fractures in each layer of the reservoir, which is conducive to a comprehensive understanding of the fracture structure inside the reservoir. It can also help technicians evaluate the fracturing effect from a spatial perspective and provide a basis for the formulation and optimization of fracturing construction schemes. Attached Figure Description

[0089] Figure 1 This is a flowchart of a three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data according to the present invention.

[0090] Figure 2 This is a schematic diagram of the structure of the fracturing fracture distribution prediction model of the present invention.

[0091] Figure 3 This is a schematic diagram of the multi-source data dynamic weighted fusion module of the present invention.

[0092] Figure 4 This is a three-dimensional fracture distribution reconstruction diagram of the second fracturing stage of the present invention.

[0093] Figure 5 This is a reconstruction of the three-dimensional fracture distribution in the third fracturing stage.

[0094] Figure 6 This is a three-dimensional fracture distribution reconstruction diagram of the fifth fracturing stage of the present invention. Detailed Implementation

[0095] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0096] This embodiment discloses a three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data, such as... Figure 1 As shown, it includes the following steps:

[0097] Step 1: Establish a fracturing model based on on-site fracturing data. Use the fracturing model to simulate the fracturing process and microseismic events in horizontal wells within the reservoir. Obtain reservoir attribute datasets, pumping design datasets, wellbore-natural fracture distribution image sets, fracturing fracture distribution image sets, and microseismic event datasets. Construct a multi-source database for fracturing fracture characterization. This includes the following sub-steps:

[0098] Step 1.1: Obtain multiple fracturing construction site data, including reservoir data, fracturing design schemes, and well completion data. Based on the fracturing construction site data, including reservoir data, fracturing design schemes, and well completion data, construct a reservoir attribute dataset, a pumping design dataset, and a wellbore-natural fracture distribution image set.

[0099] The fracturing operation site data includes reservoir data, fracturing design schemes, and well completion data. The reservoir data includes multiple reservoir physical property parameters, namely porosity, permeability, water saturation, reservoir pressure, triaxial principal stress, Young's modulus, and Poisson's ratio, as shown in Table 1. Each reservoir physical property parameter is saved in CSV format to obtain multiple reservoir attribute data and construct a reservoir attribute dataset.

[0100] Table 1. Reservoir physical property characteristics.

[0101] .

[0102] The fracturing design scheme includes multiple fracturing construction characteristic parameters, namely, pumping rate, fracturing fluid type, proppant type and proppant concentration between fracturing horizontal well sections. Each fracturing construction characteristic parameter is saved in CSV format to obtain multiple pumping design data and construct a pumping design dataset.

[0103] The well completion data includes wellbore perforation images and natural fracture distribution images of each fractured section in the fractured horizontal well. The natural fracture distribution image is a two-dimensional spatial distribution determined based on geological modeling, and together with the wellbore perforation image generated based on the perforation location, a wellbore-natural fracture image is determined. The wellbore perforation images and natural fracture distribution images of each fractured section in the fractured horizontal well are all saved in BMP format to obtain multiple wellbore-natural fracture images, and a wellbore-natural fracture distribution image set is constructed.

[0104] Step 1.2: Based on the data from each fracturing construction site, construct fracturing models in the fracture propagation numerical simulator to obtain multiple fracturing models under different geological conditions and fracturing design schemes. Perform fracturing construction simulation for each fracturing model to generate fracturing fractures and obtain fracturing fracture distribution images corresponding to different geological conditions and fracturing design schemes. Construct a fracturing fracture distribution image set.

[0105] Furthermore, the fracturing model is constructed based on the reservoir physical property parameters and fracturing design scheme in step 1.1, resulting in a total of 50 fracturing models with different geological conditions and fracturing design schemes. In this embodiment, the model size of the fracturing model is set to 600m × 400m × 20m, and the fracturing model contains multiple grid cells with a size of 10m × 10m × 2m, covering the horizontal well section in the reservoir and the surrounding reservoir stimulation volume range; the fracturing model is vertically divided into four layers, each with a thickness of 5m, and the reservoir physical properties of each layer are set according to the reservoir data; the fracturing model includes a horizontal well model, the wellbore of which is set along the length of the fracturing model, and multiple fracturing segments are set in the horizontal well model according to the fracturing design scheme, with the fracturing segments connected... The spacing is set to 80~120m. The perforation cluster positions of the horizontal well model are set according to the wellbore perforation images in the well completion data. Multiple perforation clusters are set in each fracturing section, and the cluster spacing between each perforation cluster is set to 15~25m. The wellbore perforation images and natural fracture distribution images of each fracturing section in the well completion data are imported into the fracturing model as the initial fractures of the fracturing model. The fracturing fracture distribution images of the fracturing model constructed under different geological conditions and fracturing design schemes are simulated using a fracture propagation numerical simulator. A set of fracturing fracture distribution images covering different geological conditions and fracturing construction design schemes is obtained.

[0106] Step 1.3: When simulating fracturing operations based on each fracturing model, the spatial location of the fracture mesh elements and the shear displacement on the fracture mesh elements in each fracturing model are obtained. The seismic moment is determined, and microseismic events are generated using a preset microseismic data generation program. This yields microseismic event data including the three-dimensional spatial location and magnitude of the microseismic events. In this embodiment, the shear modulus is set to... Pa, minimum magnitude is -3, area of ​​minimum connected component is 0.1 The microseismic event data is saved as a CSV file to create a microseismic event dataset.

[0107] Step 1.4: Construct a multi-source database for characterizing fracturing fractures based on reservoir attribute datasets, pumping design datasets, wellbore-natural fracture distribution image sets, fracturing fracture distribution image sets, and microseismic event datasets.

[0108] Step 2 involves preprocessing the multi-source data from the multi-source database for fracturing fracture characterization to obtain two-dimensional microseismic feature maps of each reservoir layer and grayscale images of the wellbore-natural fracture distribution. This includes the following sub-steps:

[0109] Step 2.1: Normalize the reservoir physical property parameters in the reservoir attribute dataset and the fracturing operation characteristic parameters in the pumping design dataset.

[0110] The normalization calculation formula is as follows:

[0111] ;

[0112] In the formula, These are the normalized parameters; These are unnormalized parameters; The minimum value of the parameter; This represents the maximum value of the parameter.

[0113] Step 2.2: Preprocess the image data in the wellbore-natural fracture distribution image set and the hydraulic fracture distribution image set. For each wellbore hole image, each natural fracture distribution image and each hydraulic fracture distribution image, first perform binarization processing according to the preset grayscale threshold, then perform grayscale mapping and crop to the specified size to obtain the wellbore hole grayscale image, the natural fracture distribution grayscale image and the hydraulic fracture distribution grayscale image, and use the hydraulic fracture distribution grayscale image as the label.

[0114] Specifically, in this embodiment, when binarizing the wellbore hole image, the natural fracture distribution image, and the hydraulic fracture distribution image, the grayscale threshold of the wellbore hole image and the natural fracture distribution image is preset to 35. The portion of the wellbore hole image and the natural fracture distribution image with a grayscale value less than the preset grayscale threshold of 35 is used as the background grayscale and assigned a value of 0, while the portion of the wellbore hole image and the natural fracture distribution image with a grayscale value not less than the preset grayscale threshold of 35 is assigned a value of 1. The grayscale threshold of the hydraulic fracture distribution image is preset to 30. The portion of the hydraulic fracture distribution image with a grayscale value less than the preset grayscale threshold of 30 is used as the background grayscale and assigned a value of 0, while the portion of the hydraulic fracture distribution image with a grayscale value not less than the preset grayscale threshold of 30 is assigned a value of 1.

[0115] When performing grayscale mapping on the binarized images of each wellbore hole, each natural fracture distribution image, and each hydraulic fracture distribution image, an identifier mask is used to set the grayscale value of each region in the image. The grayscale value of the hydraulic fracture region in each image is set to 1, the grayscale value of the natural fracture region is set to 2, and the grayscale value of the wellbore hole region is set to 3.

[0116] Finally, the dimensions of each wellbore hole image, each natural fracture distribution image, and each hydraulic fracture distribution image were all cropped to [size missing]. Each hydraulic fracturing fracture distribution image is stored separately as a label.

[0117] Step 2.3: Preprocess the microseismic event data in the microseismic event dataset.

[0118] For each microseismic event data point, translation, rotation, and scaling transformations are performed based on the microseismic 3D spatial location of the data to unify the microseismic 3D spatial location to the coordinate system of the fracturing model. After dividing the microseismic event data into layers based on reservoir properties and the layer thickness of the fracture image (preset layer thickness), a two-dimensional spatial grid with the same size as the fracturing fracture distribution image is constructed. The size of the two-dimensional spatial grid is... Furthermore, the two-dimensional spatial grid includes three types of feature channels: microseismic event density channel, magnitude-weighted energy channel, and... Value distribution channel, used to extract microseismic event density, magnitude-weighted energy, and... Value distribution.

[0119] The microseismic event density channel is used to describe the spatial concentration of microseismic events within a plane, and is specifically configured as follows:

[0120] ;

[0121] in,

[0122] ;

[0123] In the formula, For the first Microseismic event density of layers in a two-dimensional spatial grid; This refers to the sequence number of the microseismic event; This represents the total number of microseismic events. It is a two-dimensional Gaussian kernel function; The x-coordinate of the two-dimensional spatial grid; The ordinate of the two-dimensional spatial grid; Two-dimensional spatial grid Place and No. Micro-earthquake events in The difference in the index of the direction; Two-dimensional spatial grid Place and No. Micro-earthquake events in The difference in the index of the direction; is the base of the natural exponential function; For the first Horizontal index of microseismic events in a two-dimensional spatial grid; For the first Vertical index of microseismic events in a two-dimensional spatial grid; This is the spatial smoothing coefficient, with a value ranging from 2 to 5.

[0124] The magnitude-weighted energy channel is used to reflect the energy differences of microseismic events in different regions, and is specifically configured as follows:

[0125] ;

[0126] In the formula, For the first Magnitude-weighted energy of the layer; This is the magnitude-energy weighting coefficient, with a value ranging from 1.0 to 1.5. When, it is expressed as linear weighting, emphasizing the number of microseismic events, when When expressed as nonlinear weighting, it emphasizes the energy release of microseismic events; For the first The magnitude of a micro-earthquake event.

[0127] To quantify the complexity of microseismic events in different regions, a local sliding window is set to determine the... The value distribution channel is:

[0128] ;

[0129] In the formula, These are local frequency-magnitude distribution parameters; The average magnitude of microseismic events within the local sliding window; The minimum magnitude within the local sliding window; The intervals are for classifying earthquake magnitudes.

[0130] The The value is used to indicate the degree of fracture activity, and the numerical range is typically 0.5 to 1.5. The smaller the value, the higher the proportion of large earthquakes and the stronger the crack activity. The larger the value, the more dispersed the energy and the more fragmented the cracks become.

[0131] For each layer, the spatial distribution of microseismic events is used as input to a two-dimensional spatial grid. A Gaussian kernel density estimation method is employed to generate continuous rasterized feature images to characterize the spatial clustering of microseismic events, and to extract microseismic event density, magnitude-weighted energy, and... The value distribution was normalized to limit the numerical range of microseismic event density and magnitude-weighted energy. The dimensions at each layer are obtained as follows: Two-dimensional microseismic feature map.

[0132] Step 2.4: The grayscale images of wellbore perforations and natural fracture distribution at each layer are stitched together to generate a dimension-... The image contains a grayscale image of the distribution of natural fractures in the wellbore, which has two feature channels: a natural fracture channel and a wellbore borehole location channel, used to extract the locations of natural fractures and wellbore boreholes.

[0133] Step 3: Construct a fracturing fracture distribution prediction model based on multi-source data feature fusion, such as... Figure 2 As shown, the fracturing fracture distribution prediction model includes a reservoir attribute data processing module, a pumping design data processing module, a microseismic event data processing module, an image convolution module, a multi-source data dynamic weighted fusion module, and a fracture distribution prediction module.

[0134] Furthermore, the reservoir attribute data processing module is used to process reservoir attribute data, first processing the input reservoir attribute data... Dimensions converted ,in, , , , The dimensions of the reservoir attribute data are represented by the number of channels, depth, height, and width. Then, multi-layer convolution processing is applied to the dimensionally transformed reservoir attribute data. Convolutional layers extract the basic features of reservoir attribute data. After ReLU activation layers enhance the nonlinear expressive power, the data is then processed sequentially through three identical layers. After the convolutional layer extracts deep features, it utilizes... The convolutional layer performs channel compression, and the output dimension is... The reservoir properties and characteristics.

[0135] The pumping design data processing module is used to process pumping design data. First, it converts the input pumping design data to frequency domain data using a Fast Fourier Transform. Then, it sequentially inputs the frequency domain data into three identical... After extracting local spectral features from the convolutional layer, then... The convolutional layer performs feature integration and compression, and the output dimension is... The pumping embedding feature.

[0136] The microseismic event data processing module is used to process microseismic event data, first obtaining the three-channel feature tensor from the input two-dimensional microseismic feature map. For each channel feature tensor, the channel feature tensors are input sequentially into... After extracting local features from convolutional layers and ReLU activation layers, then... The convolutional layer performs channel compression, and the output dimension is... Microseismic embedding features .

[0137] The image convolution module is used to process grayscale images of wellbore-natural fracture distribution. It includes three sequentially connected image convolution units with identical structures. Each image convolution unit consists of a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The first image convolution unit uses a convolutional layer with 2 input channels and 32 output channels. Convolutional layers are used to extract local spatial features; the second image convolutional unit uses 32 input channels and 32 output channels. Convolutional layers are used to extract deep semantic features; the third image convolutional unit uses 32 input channels and 2 output channels. Convolutional layers are used for channel compression while preserving spatial resolution, with an output dimension of [missing information]. Image features.

[0138] The multi-source data dynamic weighted fusion module is used to obtain the multi-source fusion feature tensor, such as... Figure 3 As shown, it includes a splicing layer, a max pooling layer and an average pooling layer arranged side by side, and layers arranged sequentially. Convolutional layer, ReLU activation layer, Convolutional layers and sigmoid activation layers, by introducing an attention mechanism, automatically learn the importance weights of various features in crack prediction, thereby achieving dynamic weighted fusion of multi-source input features.

[0139] The fracture distribution prediction module is used to predict the fracture distribution in the reservoir. It consists of multiple parallel and identical improved U-Net structures. The number of improved U-Net structures is set according to the number of layers. Each improved U-Net structure includes an encoder, a decoder, a channel attention module, and an output layer. The encoder is used for multi-scale spatial feature extraction and downsampling compression, and includes multiple coding layers. The number of channels in each coding layer increases progressively to enhance the expression of deep semantic information. The coding layers are used for multi-scale spatial feature extraction and downsampling compression, and include sequentially connected... Convolutional layers, batch normalization layers, ReLU activation layers, and layers using... The decoder employs a window-downsampling max-pooling layer, where a ReLU activation layer enhances non-linear expressive power, and the max-pooling layer achieves feature compression and main feature extraction. The decoder utilizes a hierarchical design combining upsampling, convolution processing, batch normalization, ReLU activation, and skip connections to recover spatial resolution and detail information layer by layer. It includes multiple decoding layers, each containing two structurally identical and sequentially connected decoding convolutional units. Each decoding convolutional unit includes sequentially connected... Convolutional layers, ReLU activation layers and The convolutional layer, through upsampling and skip connections, recovers spatial resolution layer by layer, restoring spatial resolution and detail information. The channel attention module runs through both the encoder and decoder processes, enhancing the model's ability to identify crack development regions. The output layer is equipped with... Convolutional layers and sigmoid activation layers are used to output a crack distribution probability map, where the value range of each pixel in the crack distribution probability map is [range missing]. , is used to represent the probability of crack development at a pixel.

[0140] Specifically, when the fracturing fracture distribution prediction model is applied to fracture distribution prediction, it includes the following sub-steps:

[0141] Step 3.1: The reservoir attribute features of each reservoir attribute data in the reservoir attribute dataset are extracted using the reservoir attribute data processing module. Then, the pumping embedding features of each pumping design data in the pumping design dataset are extracted using the pumping design data processing module. The extracted reservoir attribute features and pumping embedding features are then summed element-wise using an element-wise addition method, and bilinear interpolation is used to upsample to... Resolution, resulting in dimensions The fusion feature map contains reservoir-pump fusion features of each layer and is provided with reservoir-pump fusion feature channels.

[0142] Step 3.2: Use the image convolution module to extract the image features of each wellbore-natural fracture distribution grayscale image in the wellbore-natural fracture distribution image set, and use the microseismic event data processing module to extract the microseismic embedding features of each microseismic event data in the microseismic event dataset.

[0143] Step 3.3: The extracted fusion feature map, microseismic embedding features, and image features are fused using the multi-source data dynamic weighted fusion module to obtain the multi-source fusion feature tensor. The fracture distribution image is used as the label for supervised learning. The specific process is as follows:

[0144] First, the fused feature map, microseismic embedding features, and image features input to the multi-source data dynamic weighted fusion module are stitched together to obtain a dimension of The input feature data has six channels corresponding to the reservoir-pump fusion feature channel, the natural fracture channel, the wellbore location channel, the microseismic event density channel, the magnitude-weighted energy channel, and the... Value distribution channel.

[0145] Channel attention is then applied to the input features, which are then fed into average pooling and max pooling layers, respectively. After average pooling and max pooling, the input features are fed into sequentially connected layers. Convolutional layer, ReLU activation layer, Feature mapping is performed in the convolutional layer to obtain average pooling and max pooling features. These features are then fused and input into a sigmoid function layer for normalization, yielding channel weight coefficients. These channel weight coefficients are then multiplied and weighted channel-by-channel by the input features to obtain a weighted sum of the resulting features. The multi-source fusion feature tensor is used, and the fracturing crack distribution image is used as the label for supervised learning.

[0146] Step 3.4: Input the multi-source fusion feature tensor into the fracture distribution prediction module, and use the fracture distribution prediction module to extract spatial features from the multi-source fusion feature tensor to predict the fracture distribution in the reservoir. This specifically includes the following sub-steps:

[0147] Step 3.4.1: To ensure that the dimension and computational accuracy of the input multi-source fusion feature tensor match, the multi-source fusion feature tensor is normalized and then input into the encoder of the improved U-Net structure. The encoder processes the input and outputs a set of feature matrices. , ,in, , , , The feature matrices extracted for each structural layer of the encoder, where the feature matrices are... The dimension is Feature matrix The dimension is Feature matrix The dimension is Feature matrix The dimension is .

[0148] Step 3.4.2: Calculate the encoder output feature matrix set respectively. The channel attention weights of each feature matrix are calculated using the following formula:

[0149] ;

[0150] In the formula, For the set of characteristic matrices The index of the feature matrix; This is the weighted feature matrix; For the first The feature matrix output by each coding layer; For the first Channel weight matrix of each feature matrix; Weighted operation for the channel.

[0151] Step 3.4.3: The set of feature matrices after channel weighting and the set of feature matrices output by the encoder are used together as the input features of the decoder. The decoder is then used to process the set of feature matrices. Decode each feature matrix in the encoder, using the feature matrix output from the deepest layer. As input features for the decoder, each decoding layer first upsamples the input features using bilinear interpolation, and then compares them with the feature matrix output by the corresponding coding layer. Skip connections and splicing are performed in the channel dimension, followed by refinement and channel adjustment by the decoding convolutional units within the decoding layer. Finally, the decoded features are obtained by decoding layer by layer in the decoder in order from deep to shallow. .

[0152] In this embodiment, the feature matrices obtained from each decoding layer of the decoder are obtained in descending order of depth. , , , , where the characteristic matrix The dimension is Feature matrix The dimension is Feature matrix The dimension is Feature matrix The dimension is Furthermore, bilinear interpolation is used for upsampling at each decoding layer to ensure smoothness and spatial consistency. The upsampled features and the corresponding encoder outputs are then compared. By performing jump connections and splicing, high-level semantics and low-level details can be effectively integrated. The spliced ​​features are then processed by convolution and normalization and input into the next decoding layer to form a progressively refined crack space feature representation.

[0153] Step 3.4.4: Decode the features obtained by the decoder. Input to the output layer, via After the convolutional layer compresses its channels, it is input into the Sigmoid activation layer, generating a dimension of A crack distribution probability map, wherein the value range of each pixel on the crack distribution probability map is... , is used to represent the probability of crack development at a pixel, where the closer the pixel value is to 1, the higher the degree of crack development, and the closer it is to 0, the non-crack area.

[0154] Step 4: Train and validate the fracturing fracture distribution prediction model based on multi-source data feature fusion using multi-source data from the fracturing fracture characterization multi-source database to obtain the validated fracturing fracture distribution prediction model. This step includes the following sub-steps:

[0155] Step 4.1: Randomly allocate various types of data from each dataset in the multi-source database for fracturing fracture characterization to the training set and validation set according to a preset ratio. In this embodiment, the ratio of the amount of data in the training set and the validation set of the same type of data is 8:2.

[0156] Step 4.2: Set the initial learning rate of the fracturing fracture distribution prediction model to 0.0001 and the accuracy value to [value missing]. The training process employs the Adam optimization algorithm, and the model gradually converges through a dynamic learning rate scheduling strategy.

[0157] Step 4.3: Input various types of data from the training set into the fracturing fracture distribution prediction model. Use the training set to train the fracturing fracture distribution prediction model to predict the fracture distribution in the reservoir, and obtain the comprehensive loss function value of the fracturing fracture distribution prediction model. for:

[0158] ;

[0159] In the formula, , All are loss weighting coefficients; For focus loss function; for The coefficient is used to measure the degree of overlap between the predicted results and the actual crack area. The labeled image is the distribution image of the hydraulic fractures, which serves as the label for the multi-source fusion feature tensor. The results are the predictions from the fracturing fracture distribution prediction model.

[0160] Step 4.4: Compare the comprehensive loss function value of the fracturing fracture distribution prediction model with the preset accuracy value. If the comprehensive loss function value has reached the preset accuracy value, the training of the fracturing fracture distribution prediction model ends; otherwise, the learning rate of the fracturing fracture distribution prediction model is adjusted, and the process returns to step 4.3.

[0161] Step 4.5: Input various types of data from the validation set into the fracturing fracture distribution prediction model, validate the fracturing fracture distribution prediction model using the validation set, obtain the comprehensive loss function value of the fracturing fracture distribution prediction model, and compare it with the preset accuracy value. If the comprehensive loss function value has reached the preset accuracy value, the validated fracturing fracture distribution prediction model is obtained; otherwise, adjust the learning rate of the fracturing fracture distribution prediction model and return to step 4.3 to continue training the fracturing fracture distribution prediction model.

[0162] Step 4.6: Obtain the validated fracturing fracture distribution prediction model.

[0163] In this embodiment, during the training of the fracturing fracture distribution prediction model, the channel attention module dynamically adjusts the weights based on the correlation between features from different data sources, achieving adaptive weighted fusion of reservoir features, pumping features, and microseismic features. Through forward propagation and backward gradient updates, the fracture distribution prediction module is continuously optimized until the overall loss of the module stabilizes. At this point, the module is considered to have converged, resulting in the trained fracturing fracture distribution prediction model.

[0164] Step 5: Based on the reservoir attribute data and pumping design data collected on site, input the grayscale image of wellbore-natural fracture distribution constructed by the on-site geological data and perforation scheme, as well as the simulated microseismic event data, into the validated fracturing fracture distribution prediction model. Use the fracturing fracture distribution prediction model to predict the fracturing fracture distribution and generate a three-dimensional fracture distribution reconstruction map.

[0165] In this embodiment, the depth of the horizontal well model in the fracturing model was 3570 m and the vertical depth was 2410 m. The maximum horizontal principal stress of the reservoir ranged from 78 to 86 MPa, the minimum horizontal principal stress ranged from 66 to 70 MPa, the vertical stress ranged from 73 to 77 MPa, the formation pressure ranged from 33 to 46 MPa, the Young's modulus ranged from 28 to 33 GPa, and the Poisson's ratio ranged from 0.23 to 0.27. The reservoir porosity was relatively high and stable, with logging porosity values ​​of 5.8% to 6.4% and an average porosity of 6.1%. The total number of grids was 24,000, and the size of each grid cell was 10m × 10m × 2m. Three fracturing sections of the well were selected, and the three-dimensional fracturing fracture distribution prediction effect was verified using the validated fracturing fracture distribution prediction model. The fracturing construction characteristic parameters of each fracturing section are shown in Table 2.

[0166] Table 2. Characteristic parameters of fracturing operation in the three fracturing stages.

[0167] .

[0168] The size was predicted using a fracturing fracture distribution prediction model. The crack distribution probability map, wherein the value range of each pixel in the crack distribution probability map is... The value of each pixel represents the probability of crack development at that location.

[0169] The crack distribution probability map is divided into non-crack areas and crack development areas according to a preset threshold of 0.35. Pixels with a probability value less than 0.35 are identified as non-crack areas and their pixel values ​​are set to 0. Pixels with a probability value not less than 0.35 are identified as crack development areas and their pixel values ​​are set to 1, thereby generating a crack prediction image.

[0170] A bilinear interpolation algorithm was used to spatially register the generated fracture prediction images with the original formation model of the actual reservoir, aligning the fracture prediction images with the grid of the original geological model. Fracture prediction images of each layer were then stacked along the vertical depth direction of the actual reservoir to obtain a three-dimensional fracture distribution reconstruction map of each fracturing segment, as shown below. Figures 4-6 As shown, this enables accurate prediction and intuitive display of the fracture conditions inside the reservoir.

[0171] In summary, the three-dimensional characterization method for fracturing fracture distribution based on intelligent fusion of multi-source data proposed in this invention has achieved innovative breakthroughs in multi-source data feature fusion, reservoir physical constraint feature extraction, and fracturing fracture distribution prediction model construction, realizing high-precision, interpretable, and visual prediction of fracturing fracture distribution. This method not only improves the accuracy and robustness of reservoir fracture distribution prediction but also provides a basis for optimizing fracturing construction design, evaluating reservoir stimulation effects, and predicting oil and gas production capacity, demonstrating significant engineering application value and promising prospects for wider application.

[0172] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for three-dimensional characterization of fracture distribution based on intelligent fusion of multi-source data, characterized in that, Includes the following steps: Step 1: Establish a fracturing model based on the fracturing construction site data, and use the fracturing model to simulate the fracturing construction process and microseismic events of horizontal wells in the reservoir. Obtain reservoir attribute datasets, pumping design datasets, wellbore-natural fracture distribution image sets, fracturing fracture distribution image sets, and microseismic event datasets to construct a multi-source database for fracturing fracture characterization. Step 2: Preprocess the multi-source data in the multi-source database for fracturing fracture characterization to obtain two-dimensional microseismic feature maps of each reservoir layer and grayscale images of wellbore-natural fracture distribution. Step 3: Construct a fracturing fracture distribution prediction model based on multi-source data feature fusion. The fracturing fracture distribution prediction model includes a reservoir attribute data processing module, a pumping design data processing module, a microseismic event data processing module, an image convolution module, a multi-source data dynamic weighted fusion module, and a fracture distribution prediction module. Step 4: Use multi-source data from the multi-source database of fracturing fracture characterization to train and validate the fracturing fracture distribution prediction model based on multi-source data feature fusion, and obtain the validated fracturing fracture distribution prediction model. Step 5: Based on the reservoir attribute data and pumping design data collected on site, input the wellbore-natural fracture distribution grayscale image constructed based on the on-site geological data and perforation scheme and the simulated microseismic event data into the validated fracturing fracture distribution prediction model. Use the fracturing fracture distribution prediction model to predict the fracturing fracture distribution and generate a three-dimensional fracture distribution reconstruction map. The microseismic event data processing module is used to process microseismic event data, first obtaining the three-channel feature tensor from the input two-dimensional microseismic feature map. For each channel feature tensor, the channel feature tensors are input sequentially into... After extracting local features from convolutional layers and ReLU activation layers, then... Convolutional layers perform channel compression to output microseismic embedding features. ; The image convolution module is used to process grayscale images of wellbore-natural fracture distribution. It includes three sequentially connected image convolution units with identical structures. Each image convolution unit consists of a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The first image convolution unit uses a convolutional layer with 2 input channels and 32 output channels. Convolutional layers are used to extract local spatial features; the second image convolutional unit uses 32 input channels and 32 output channels. Convolutional layers are used to extract deep semantic features; the third... Image convolutional units use 32 input channels and 2 output channels. Convolutional layers are used for channel compression and outputting image features.

2. The method for three-dimensional characterization of fracturing fracture distribution based on intelligent fusion of multi-source data according to claim 1, characterized in that, Step 1 includes the following sub-steps: Step 1.1: Obtain multiple fracturing operation site data, including reservoir data, fracturing design schemes and well completion data, and construct a reservoir attribute dataset, a pumping design dataset and a wellbore-natural fracture distribution image set; The reservoir data includes multiple reservoir physical property parameters, namely porosity, permeability, water saturation, reservoir pressure, triaxial principal stress, Young's modulus, and Poisson's ratio; the fracturing design scheme includes multiple fracturing construction characteristic parameters, namely pumping rate between fracturing horizontal well sections, fracturing fluid type, proppant type, and proppant concentration; the completion data includes wellbore perforation images and natural fracture distribution images of each fracturing section in the fracturing horizontal well, used to determine the wellbore-natural fracture image; Step 1.2: Construct fracturing models in the fracture propagation numerical simulator based on the data from each fracturing construction site to obtain multiple fracturing models under different geological conditions and fracturing design schemes. Perform fracturing construction simulation for each fracturing model to generate fracturing fractures, obtain fracturing fracture distribution images, and construct a fracturing fracture distribution image set. Step 1.3: When simulating fracturing construction based on each fracturing model, obtain the spatial position of the fracture mesh element and the shear displacement on the fracture mesh element in each fracturing model, determine the seismic moment, and use the preset microseismic data generation program to generate microseismic events, obtain microseismic event data including the three-dimensional spatial position and magnitude of microseismic events, and establish a microseismic event dataset. Step 1.4: Construct a multi-source database for characterizing fracturing fractures based on reservoir attribute datasets, pumping design datasets, wellbore-natural fracture distribution image sets, fracturing fracture distribution image sets, and microseismic event datasets.

3. The method for three-dimensional characterization of fracturing fracture distribution based on intelligent fusion of multi-source data according to claim 2, characterized in that, In step 1.3, a shear displacement threshold is set based on the minimum magnitude, shear modulus, and minimum fracture zone area in the fracturing model. for: ； In the formula, It is the seismic moment; Shear modulus; The minimum area of ​​the fractured region; Using a breadth-first search algorithm, models with shear displacements greater than a shear displacement threshold are selected from the fracturing model. Given connected components, calculate the area of ​​each connected component. centroid location and mean shear displacement ; For each connected block, calculate the seismic moment of the connected block based on its shear displacement and area. Moment and Magnitude The calculation formula is: ； ； In the formula, This represents the average shear displacement. It is a logarithmic function with base 10; The area, centroid location, average shear displacement, seismic moment, and moment magnitude of each connected block are saved to the crack propagation mesh file to form multiple microseismic event data and construct a microseismic event dataset.

4. The method for three-dimensional characterization of fracturing fracture distribution based on intelligent fusion of multi-source data according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step 2.1: Normalize the reservoir physical property parameters in the reservoir attribute dataset and the fracturing operation characteristic parameters in the pumping design dataset. Step 2.2: Preprocess the image data in the wellbore-natural fracture distribution image set and the hydraulic fracture distribution image set. For each wellbore hole image, each natural fracture distribution image and each hydraulic fracture distribution image, first perform binarization according to the preset grayscale threshold, then perform grayscale mapping, and then crop to the specified size to obtain the wellbore hole grayscale image, the natural fracture distribution grayscale image and the hydraulic fracture distribution grayscale image, and use the hydraulic fracture distribution grayscale image as the label; Step 2.3: Preprocess the microseismic event data in the microseismic event dataset; For each microseismic event data point, after spatial transformation and layer division based on the three-dimensional spatial location of the microseismic events, a two-dimensional spatial grid with the same size as the fracturing fracture distribution image is constructed. This two-dimensional spatial grid includes three types of feature channels: microseismic event density channel, magnitude-weighted energy channel, and... Value distribution channel, used to extract microseismic event density, magnitude-weighted energy, and... Value distribution; For each layer, using the spatial distribution of microseismic events as input to a two-dimensional spatial grid, a continuous rasterized feature image is generated using the Gaussian kernel density estimation method. Microseismic event density, magnitude-weighted energy, and other parameters are then extracted. The values ​​are distributed and normalized to limit the range of microseismic event density and magnitude-weighted energy, resulting in two-dimensional microseismic feature maps at each layer. Step 2.4: The grayscale images of wellbore openings and natural fracture distribution at each layer are stitched together to generate a grayscale image of wellbore-natural fracture distribution. The grayscale image of wellbore-natural fracture distribution has two feature channels, namely the natural fracture channel and the wellbore opening location channel, which are used to extract the locations of natural fractures and wellbore openings.

5. The method for three-dimensional characterization of fracturing fracture distribution based on intelligent fusion of multi-source data according to claim 4, characterized in that, In step 2, the microseismic event density channel is set as follows: ； in, ； In the formula, For the first Microseismic event density of layers in a two-dimensional spatial grid; This refers to the sequence number of the microseismic event; This represents the total number of microseismic events. It is a two-dimensional Gaussian kernel function; The x-coordinate of the two-dimensional spatial grid; The ordinate of the two-dimensional spatial grid; Two-dimensional spatial grid Place and No. Micro-earthquake events in The difference in the index of the direction; Two-dimensional spatial grid Place and No. Micro-earthquake events in The difference in the index of the direction; is the base of the natural exponential function; For the first Horizontal index of microseismic events in a two-dimensional spatial grid; For the first Vertical index of microseismic events in a two-dimensional spatial grid; It is the spatial smoothing coefficient; The magnitude-weighted energy channel is configured as follows: ； In the formula, For the first Magnitude-weighted energy of the layer; This is the magnitude-energy weighting coefficient; For the first The magnitude of a micro-earthquake event; Set a partial sliding window to determine the... The value distribution channel is: ； In the formula, These are local frequency-magnitude distribution parameters; The average magnitude of microseismic events within the local sliding window; The minimum magnitude within the local sliding window; The intervals are for classifying earthquake magnitudes.

6. The method for three-dimensional characterization of fracturing fracture distribution based on intelligent fusion of multi-source data according to claim 1, characterized in that, The reservoir attribute data processing module is used to process reservoir attribute data, first processing the input reservoir attribute data... Dimensions converted ,in, , , , The dimensions of the reservoir attribute data are represented by the number of channels, depth, height, and width. Then, multi-layer convolution processing is applied to the dimensionally transformed reservoir attribute data. Convolutional layers extract the basic features of reservoir attribute data. After ReLU activation layers enhance the nonlinear expressive power, the data is then processed sequentially through three identical layers. After the convolutional layer extracts deep features, it utilizes... The convolutional layer performs channel compression and outputs reservoir attribute features. The pumping design data processing module is used to process pumping design data. First, it converts the input pumping design data to frequency domain data using a Fast Fourier Transform. Then, it sequentially inputs the frequency domain data into three identical... After extracting local spectral features from the convolutional layer, then... Convolutional layers perform feature integration and compression, outputting pumped embedded features; The multi-source data dynamic weighted fusion module is used to obtain multi-source fusion feature tensors, including a splicing layer, a parallel max pooling layer and an average pooling layer, and sequentially set... Convolutional layer, ReLU activation layer, Convolutional layers and sigmoid activation layers automatically learn the importance weights of various features in crack prediction by introducing an attention mechanism; The fracture distribution prediction module is used to predict the fracture distribution in the reservoir. It consists of multiple parallel and identical improved U-Net structures. The number of improved U-Net structures is set according to the number of layers. Each improved U-Net structure includes an encoder, a decoder, a channel attention module, and an output layer. The encoder includes multiple encoding layers, each of which comprises sequentially connected... Convolutional layers, batch normalization layers, ReLU activation layers, and layers using... A window-downsampled max-pooling layer; the decoder includes multiple decoding layers, each decoding layer containing two identical and sequentially connected decoding convolutional units, the decoding convolutional unit including... Convolutional layers, ReLU activation layers and The convolutional layer, through upsampling and skip connections, recovers spatial resolution layer by layer, used to restore spatial resolution and detail information; the channel attention module runs through the encoder's encoding process and the decoder's decoding process, used to enhance the model's ability to identify crack development regions; the output layer is equipped with... Convolutional layers and sigmoid activation layers are used to output a crack distribution probability map.

7. The method for three-dimensional characterization of fracturing fracture distribution based on intelligent fusion of multi-source data according to claim 6, characterized in that, When the fracturing fracture distribution prediction model is applied to fracture distribution prediction, it includes the following sub-steps: Step 3.1: Extract reservoir attribute features from each reservoir attribute data in the reservoir attribute dataset using the reservoir attribute data processing module, and then extract pumping embedding features from each pumping design data in the pumping design dataset using the pumping design data processing module. Add the extracted reservoir attribute features and pumping embedding features element by element using the element-wise addition method, and use bilinear interpolation upsampling to obtain a fused feature map. The fused feature map is provided with a reservoir-pumping fused feature channel. Step 3.2: Use the image convolution module to extract the image features of each wellbore-natural fracture distribution grayscale image in the wellbore-natural fracture distribution image set, and use the microseismic event data processing module to extract the microseismic embedding features of each microseismic event data in the microseismic event dataset. Step 3.3: Use the multi-source data dynamic weighted fusion module to perform feature fusion on the extracted fusion feature map, microseismic embedding features and image features to obtain a multi-source fusion feature tensor labeled with the hydraulic fracture distribution image; The multi-source data dynamic weighted fusion module first concatenates the input fused feature map, microseismic embedding features, and image features to obtain the input feature quantity. Then, channel attention is applied to the input feature quantity, and the input feature quantity is input into the average pooling layer and the max pooling layer, respectively. After average pooling and max pooling processing, the input is then input into the sequentially connected layers. Convolutional layer, ReLU activation layer, Feature mapping is performed in the convolutional layer to obtain average pooling mapping features and max pooling mapping features. The two features are fused and then input into the Sigmoid function layer for normalization to obtain channel weight coefficients. The channel weight coefficients are multiplied and weighted by the input feature values ​​channel by channel to obtain the multi-source fusion feature tensor. The fracturing crack distribution image is used as the label of the multi-source fusion feature tensor. Step 3.4: Input the multi-source fusion feature tensor into the fracture distribution prediction module, and use the fracture distribution prediction module to extract the spatial features from the multi-source fusion feature tensor to predict the fracture distribution in the reservoir.

8. The method for three-dimensional characterization of fracturing fracture distribution based on intelligent fusion of multi-source data according to claim 7, characterized in that, Step 3.4 includes the following sub-steps: Step 3.4.1: After normalizing the multi-source fusion feature tensor, input it into the encoder of the improved U-Net structure. The encoder then processes the data and outputs a set of feature matrices. , ,in, , , , These are all feature matrices extracted from each coding layer of the encoder; Step 3.4.2: Calculate the encoder output feature matrix set respectively. The channel attention weights of each feature matrix are calculated using the following formula: ； In the formula, For the set of characteristic matrices The index of the feature matrix; This is the weighted feature matrix; For the first The feature matrix output by each coding layer; For the first Channel weight matrix of each feature matrix; Weighted operation for the channel; Step 3.4.3: Use the decoder to process the feature matrix set respectively. Decode each feature matrix in the encoder, using the feature matrix output from the deepest layer. As input features for the decoder, each decoding layer first upsamples the input features using bilinear interpolation, and then compares them with the feature matrix output by the corresponding coding layer. The channels are skipped and spliced, then refined and channel-adjusted by the decoding convolutional units within the decoding layer. Finally, the decoded features are obtained by decoding layer by layer in the decoder, proceeding from deep to shallow. ; Step 3.4.4: Decode the features obtained by the decoder. Input to the output layer, via After the convolutional layer compresses its channels, the data is input into the Sigmoid activation layer to generate a crack distribution probability map. The value range of each pixel in the crack distribution probability map is [value range missing]. , is used to represent the probability of crack development at a pixel.

9. The method for three-dimensional characterization of fracturing fracture distribution based on intelligent fusion of multi-source data according to claim 1, characterized in that, Step 4 includes the following sub-steps: Step 4.1: Randomly allocate various types of data from each dataset in the multi-source database for fracturing fracture characterization to the training set and validation set according to a preset ratio; Step 4.2: Set the initial learning rate and accuracy values ​​for the fracturing fracture distribution prediction model; Step 4.3: Input various types of data from the training set into the fracturing fracture distribution prediction model, use the training set to train the fracturing fracture distribution prediction model to predict the fracture distribution in the reservoir, and obtain the comprehensive loss function value of the fracturing fracture distribution prediction model. The comprehensive loss function Will loss function and The loss function is coupled and set as follows: ； In the formula, , All are loss weighting coefficients; For focus loss function; for coefficient; For label images; The prediction results are from the fracturing fracture distribution prediction model. Step 4.4: Compare the comprehensive loss function value of the fracturing fracture distribution prediction model with the preset accuracy value. If the comprehensive loss function value has reached the preset accuracy value, the training of the fracturing fracture distribution prediction model ends; otherwise, the learning rate of the fracturing fracture distribution prediction model is adjusted, and the process returns to step 4.

3. Step 4.5: Input various types of data from the validation set into the fracturing fracture distribution prediction model, validate the fracturing fracture distribution prediction model using the validation set, obtain the comprehensive loss function value of the fracturing fracture distribution prediction model, and compare it with the preset accuracy value. If the comprehensive loss function value has reached the preset accuracy value, the validated fracturing fracture distribution prediction model is obtained; otherwise, adjust the learning rate of the fracturing fracture distribution prediction model and return to step 4.3 to continue training the fracturing fracture distribution prediction model. Step 4.6: Obtain the validated fracturing fracture distribution prediction model.

10. The method for three-dimensional characterization of fracturing fracture distribution based on intelligent fusion of multi-source data according to claim 1, characterized in that, In step 5, based on the reservoir attribute data and pumping design data collected on site, the grayscale image of the wellbore-natural fracture distribution constructed based on the on-site geological data and perforation scheme and the simulated microseismic event data are input into the validated fracturing fracture distribution prediction model, and the fracturing fracture distribution prediction model is used to predict the fracture distribution probability map. According to the preset threshold, the crack distribution probability map is divided into non-crack areas and crack development areas. Pixels with probability values ​​less than the preset threshold are judged as non-crack areas and their pixel values ​​are set to 0. Pixels with probability values ​​not less than the preset threshold are judged as crack development areas and their pixel values ​​are set to 1, thus generating a crack prediction image. The bilinear interpolation algorithm was used to spatially register the generated fracture prediction images with the original formation model of the actual reservoir, and the fracture prediction images of each layer were stacked along the vertical depth direction of the actual reservoir to obtain a three-dimensional fracture distribution reconstruction map.

Images